What is a research cohort? From question to dataset

In the fall of 1948, researchers began recruiting volunteers in Framingham, Massachusetts — a town of about 28,000 people — and eventually enrolled more than 5,000 adults who had not yet developed heart disease. Every two years the volunteers came back: blood pressure, cholesterol, habits, outcomes. The Framingham Heart Study, run since 1949 by the National Heart Institute — today's NHLBI, part of the National Institutes of Health, gave medicine the phrase 'risk factor.' It also demonstrated, across seven decades and three generations of participants, the most durable unit of work in clinical research: the cohort.

A defined group, not a pile of records

A research cohort is a group of people who share a defining characteristic — a diagnosis, an exposure, a treatment start — identified in a consistent way and observed over time. In data research, the cohort is the denominator: every rate, every risk, every comparison in the study inherits its meaning from who is in the group.

That makes the definition itself the scientific act. 'Patients with bladder pain syndrome' sounds like a population; it is actually a hundred unmade decisions. Diagnosed by whom? Coded how? Confirmed how many times? Observed since when? Two careful teams answering those questions differently will assemble two different cohorts — and report two different answers — from the same database. This is why methodologists obsess over cohort definitions the way engineers obsess over interfaces: it is the part everything else depends on.

Cohorts also come in shapes. Framingham was prospective — defined first, then followed forward for decades. Most modern database studies are retrospective: the records already exist, and the cohort is reconstructed from them, which makes the definition's precision matter even more, because no one can go back and ask. A cohort can be closed, with membership fixed at day zero, or open, with people entering and leaving as they qualify. Each shape suits different questions, and mixing them up quietly distorts every rate the study reports.

From a plain question to criteria

The path from question to dataset runs through inclusion and exclusion criteria — explicit, checkable rules for who is in. Take a plain-English question: do adults with bladder pain syndrome who start treatment A visit the emergency department less often than those who start treatment B?

A workable definition might include: age 18 or older; a qualifying diagnosis code recorded at two visits at least 30 days apart; a first prescription for A or B. And exclude: any history of bladder cancer; less than a year of observable records before that first prescription.

Every line trades sensitivity against specificity. Require two diagnosis codes and you drop real patients who saw a specialist once; require only one and you sweep in rule-out visits that were never confirmed — a familiar hazard for conditions diagnosed by exclusion, as we describe in why interstitial cystitis research has been so hard. There is no neutral choice. There are only documented ones.

The clock is part of the definition

Every cohort has a clock, and most beginner mistakes are clock mistakes.

Get the clock wrong and the data will cheerfully manufacture effects. The classic is immortal time bias: classify people by something they did after day zero — filled a second prescription, survived to a follow-up visit — and you have quietly guaranteed that everyone in that group lived long enough to do it. The treatment looks protective. The calendar did all the work.

Making the definition computable

A modern cohort definition is code, not prose. The criteria above become logic over standard vocabularies — diagnosis codes, drug codes, lab codes — that a machine can execute the same way every time. That executable form is called a computable phenotype, and it earns its own discussion in how researchers find patients in messy data.

Computability changes the economics of rigor. A definition written as versioned logic can be tested, peer-reviewed, and shared; the OHDSI community routinely publishes cohort definitions that run identically against any database mapped to the OMOP common data model, which we explain in what is the OMOP CDM?. Validation becomes ordinary engineering: pull a sample, review charts, report how often the definition was right. Good studies stress-test the definition itself, too — re-running the analysis with one diagnosis code instead of two, or a 180-day washout instead of 365, to show the answer does not hinge on a single judgment call. Clinical trials run the same intellectual exercise when they register eligibility criteria at ClinicalTrials.gov — a cohort definition is eligibility criteria for a study that observes instead of intervenes.

Where the gates belong

Consent checks and privacy floors belong inside the cohort engine, not bolted on after it. The reason is mechanical: a cohort assembled first and consent-filtered later has already touched the records of people who never said yes. An engine that filters consent at assembly never does. The same goes for size. A 'cohort' of three is not a dataset; it is a description of three people — the arithmetic we walk through in k-anonymity, explained.

This is how the cohort engine at Mithras Health is built. Definitions execute only over records carrying a study-scoped, revocable consent; a result below k = 5 cannot be released — the floor is structural, not configurable; and queries run federated, in-region, returning aggregates only. Locally hosted AI helps researchers draft definitions — suggesting codes, catching gaps — but membership is decided by the deterministic engine alone, because a cohort you cannot reproduce is a cohort you cannot trust.

Framingham's lasting gift was never really its data. It was the discipline: decide exactly whom the study is about, write the decision down, and honor it for decades. A cohort is a question made precise — and the precision deserves the same gates, the same audit, and the same respect as everything built downstream of it.